If $ARGUMENTS is empty or non-numeric, analyze the current conversation. If it is a number, treat it as a request to mine that many days of historical sessions.
Current Conversation Mode
For each mistake, capture:
- What the assistant concluded
- What the actual answer turned out to be
- What information was missing
- Whether that information was discoverable
- What concrete change would prevent the same mistake
Output shape:
### Mistake: <what went wrong>
**Assistant said:** "<quoted claim>"
**Actual answer:** "<what turned out to be true>"
**Missing info:** <missing context>
**Findable?** YES / NO
**Fix:** <skill, prompt, hook, data-source, or workflow change>
After listing mistakes, propose concrete edits and ask the user which to implement.
Historical Mode
Step 1: Pre-Filter
Use a mining script or equivalent search only as a pre-filter to find candidate sessions.
Skip frozen or retry-only sessions caused by infrastructure issues rather than assistant behavior.
Step 2: Deep Read
Read flagged sessions message by message. For each user intervention, ask whether the assistant could have already done that without being told.
Step 3: Choose The Right Fix Mechanism
Map the failure to the right control surface:
| Mechanism | Best for |
|---|---|
CLAUDE.md or equivalent global instruction |
Universal behavior rules |
| Path-scoped rules | File-type-specific behavior |
| Skills | Situational knowledge |
| Memory | Persistent decisions or preferences |
| Hooks | Automated guardrails that must always run |
| Agents | Isolated delegated task types |
| Commands | Explicitly user-invoked prompts |
| Settings | Permissions, env vars, model config |
Decision heuristic:
- "Assistant keeps forgetting X" -> Hook or always-loaded rule
- "Assistant should do X when editing Y" -> Path-scoped rule
- "Assistant lacks knowledge about X during Y" -> Skill
- "Assistant should delegate X" -> Agent plus routing rule
- "I want to run X on demand" -> Command
- "Remember this decision" -> Memory
Step 4: Present Findings With Examples
Every pattern must include concrete examples from transcripts. Do not present regex counts as findings.
For each pattern include:
- What happens
- Two or three concrete examples
- Proposed fix: bucket, file, specific edit, and why it helps
Step 5: Validate Against User Experience
Check with the user that the proposed pattern matches real experience before editing prompts or skills.